Back to problems

Deploy multi-armed bandits safely

System Design · Meta · Hard

You are running a live experiment with three candidate treatments. The primary success metric is conversion, but conversions may be attributed only after some delay. A separate churn metric acts as a safety guardrail: if churn degrades beyond a predefined tolerance, the experiment must be stopped or rolled back. You plan to allocate traffic adaptively using a Thompson Sampling bandit. a) Specify the Thompson Sampling model: choose appropriate likelihoods, priors, and…

Checking your access…